I'm curious. Do you attribute this to weak and/or incomplete tests? How granular should tests be to have complete coverage so that an AI won't create a converted codebase that "passes tests" but is still functionally inaccurate?
In particular, if you put an LLM in an automated loop of "this test fails, please fix it", there is a pretty good chance that it will simply special case all of the tests, possibly in some contrived way that makes it not at all obvious when you read the code.
A truly stochastic method is more likely to hit against edge cases, rather than an agent that tends to towards idiomatic solutions and that is trained against a corpus of existing software, and burns millions of tokens/watts spinning its wheels.
^ There's plenty of business value to be found in agentic AI without reaching for it for every solution. I'd even posit agentic AI is even better when paired with focused old-fashioned squishy-brained software engineering in the loop.
Maybe a way of looking at it, to understand the nature of the issue. Have a LLM translate a novel from English to Spanish. Of course it can do that translation at speeds that no human could (score a point for AI). But how good is the Spanish translation? Is the quality better than what humans could do? Wouldn't those who are not fluent in Spanish be more easily impressed?
We then can do all kinds of configuration setups and tests, but how do we know the Spanish was translated perfectly, without a massive detail review (and being already truly bilingual in both English and Spanish)?
As is the usual case in the pursuit of perfection (which nothing in nature ever seems to be), there is going to be mistakes, costs (worth it?), and gray areas. It would be foolhardy for us not to suspect or pass it off as otherwise.
You can't. I think that's a large part of why LLMs have caught on much better with programmers: they have ways of making the computer check its own work.
Checking a document is still a laborious manual task. And completely unfulfilling.
> making the computer check its own work
Kind of like the Spanish teacher telling his students they can grade their own tests, then being surprised that Billy was always giving himself 100%, when he's nowhere near that bright or fluent.
It wouldn't be so bad, if people were more upfront with being unsure or made it clear they were extrapolating from smaller and limited data. But usually, like many of these unusually cocky LLMs, what is too often reported to the public is "perfection" and many inconvenient truths "swept underneath the carpet".
That's the culprit, because LLMs tend to forget and remove a lot of branch logic in these kinds of tasks. If unit tests don't cover these specific if/elseif/else cases, then they'll just disappear.
They'll also disappear if the LLM is allowed to modify the unit tests, because they sure like to cheat their way around into greenlit test suites. The agentic environment must disallow write access to the unit test files for the agent that writes the code.
If you implement that in your tools, you'll see quickly how the models will try to rewrite the unit tests at all cost, no matter what kind of prompting you've done. Tool policies are the only boundary to successfully guarantee this.
Source: Am building my own agentic environment because of that behavior
If LLMs can be utilized to quickly make deep testing possible, I think that's probably a net-positive.